I'm trying to deliver an AP invoice reading solution as part of larger project. I've trained a form processing model in AI Builder on 50 different invoice styles (5-10 documents each, as individual ‘collections’).
Rather than improve the accuracy with extensive training, I think I’ve just confused it. Accuracy is now less than 50% (testing is being performed using the exact same example invoices which the AI builder has been directly trained on).
I'm not sure whether it's failing to identify the correct collection for each invoice (possibly a factor of how many collections I'm using) or if it is identifying the correct collection but still failing due to the invoice format itself (i.e. small font, individual fields set close together which the model is reading as one long text string).
Suppliers in the real world don't have similar invoice layouts. I'm wondering how others have solved this issue because at the moment the model isn't even usable.
I assume I could probably improve the accuracy if I routed the invoices to individual models (specific to the supplier's invoice template), but that would mean first detecting the supplier itself, not to mention the complexity of catering for hundreds of supplier models (not even sure if it would be possible to dynamically select the model within Power Automate either).
Any advice is appreciated at this point! I've heard good things about Syntex, but understand that would still rely on the AI Builder model.
Thanks in advance!
P.S. I would have used the prebuilt invoice processing model, but I live in Australia where we use a specific (government issued) identifier for all suppliers ('ABN') which I need to pull from the invoice as part of the solution.
P.P.S. I've since found the following post, but just want to be sure that is still current advice and that there haven't been advances since then. https://powerusers.microsoft.com/t5/AI-Builder/AI-Builder-Invoice-Processing-multiple-providers-models/m-p/595153
